What should happen when a customer service AI cannot solve the problem?
In the accountability panel, Grace Dolan described separating quick customer service questions from problems that need a human response. Where that separation sits in your own service process is a design decision, and it's worth making on purpose.
List what your team can answer from an approved source and what needs investigation or judgment. Include what happens when the customer says the suggested fix didn't work, because repeating the same answer doesn't get them to a person.
Write the escalation conditions in language the service team can test. A safety concern, an account dispute or an unavailable answer may need a different route from a question about published opening hours, and the responsible team should set those routes.
The person receiving the case needs the customer's stated problem, the steps already tried, relevant records and the unresolved question. Only what's appropriate for that case should go along.
Measure resolution and repeat contact next to response speed. The objective is a useful answer or an effective route to help. A shorter automated conversation alone doesn't count.
What to try this week
Run one practice case where the customer says the suggested fix didn't work. See whether the person who receives it can pick up without asking the customer to start over.